> Each model’s responses are ranked by a high-performing judge model — typically OpenAI’s o3 — which compares outputs for quality, relevance, and clarity. These rankings are then aggregated to produce a performance score.
So there's no ground truth; they're just benchmarking how impressive an LLM's code review sounds to a different LLM. Hard to tell what to make of that.
No, they aren't. Most benchmarks use ground truth, not evaluation by another LLM. Using another LLM as verifier, aside from the obvious "quis custodiet custodes ipsos", opens an entire can of worms, such as the fact that there could be systematic biases in the evaluation. This is not in and of itself disqualifying but it should be addressed, and the article doesn't even say anything.
Even the benchmarks for maths only checked numerical answers for ground truth, which means the LLM can output a lot of nonsense and guess the correct answer to pass it
Ground truth evaluation is not that simple unless you are doing multiple-choice-style tests or something similar where the correctness of an answer can be determined by a simple process. Open ended natural language tasks like this one are incredibly difficult to evaluate and using LLMs as judge is not just the current standard, it is basically the only way to do it at scale economically.
> So there's no ground truth; they're just benchmarking how impressive an LLM's code review sounds to a different LLM. Hard to tell what to make of that.
The comment I replied to was:
> That's how 99% of 'LLM benchmark numbers' circulating on the internet work.
And that's just false. SWE-Bench verified isn't like this. Aider Polyglot isn't like this. SWE-Lancer Diamond isn't like this. The new internal benchmarks used by OpenAI in GPT-5's model card aren't like this.
Maybe this benchmark is a special snowflake and needs LLM-as-a-judge, but this doesn't invalidate the original concern: setting up a benchmark this way runs into a series of problems and is prone to show performance differences that might not be there with a different setups. Benchmarks are already hard to trust, I'm not sure how this is any more indicative than the rest.
Am I missing something? If LLM-1 is supposed to judge LLM-2, doesn't LLM-1 have to be better than LLM-2? If LLM-1 is only 40% as good at coding as LLM-2, why would you trust the LLM with the lesser knowledge?
At the heart of the P vs NP problem lies the observation that solution verification seems to be much easier than solution generation. If that applies in this context is another question but I think it is not unreasonable to assume that the judge needs to be less powerful than the performer.
Or in other words, I don't need to be a chef myself to decide if a meal is good or not.
That really doesn't hold for all problems. You can imagine any number of problems where a valid solution is easier, complexity wise, to generate than it is to validate. A trivial example is semiprime factorization. Easy to generate any semiprime, hard to factor.
Sure, it was never my intention to make it seem like a general statement, just highlighting that there is a large class of problems for which it is true.
As you point out there are many problems that higher complexity classes than NP.
Yes, especially as models are known to have a preference towards outputs of models in the same family. I suspect this leaderboard would change dramatically with different models as the judge.
There may be different/better solutions for almost all those kind of tasks. I wouldn’t be surprised if optimal answer to some of them would be refusal/defer ask, refactor first, then solve it properly.
That response is quite in line with the typical human based PR response on a first draft.
There is a possibility that machine based PR reviews are better: for instance because they are not prejudiced based on who is the initiator of the PR and because they don't take other environmental factors into account. You'd expect a machine to be more neutral, so on that front the machine should and possibly could score better. But until the models consistently outperform the humans in impartially scored quality vs a baseline of human results it is the humans that should call this, not the machines.
On something like a PR review I would. But on anything that would involve private information such as the background, gender, photographs and/or video as well as other writings by the subject I think you'd be right.
It's just that it is fairly trivial to present a PR to a machine in such a way that it can only comment on the differences in the code. I would find it surprising if that somehow led to a bias about the author. Can you give an example of how you think that would creep into such an interaction?
I'd say it's worse than that, a rubber ruler still has a definite length when not under tension etc.
This might be more like asking amateur painters to each paint a picture of a different one of the pumpkins, then judging each other's paintings without seeing the actual pumpkin that painting was based on.
It's widely accepted because it's cheap, but LLMs aren't really good judges.
It's supposed to leverage a "generate vs. critique" gap in skill level as a form of self-improvement. It's easier to judge how good food is vs. make it.
But here's the thing. When it comes to code review, you need to be effectively as skilled as the person who wrote it. There isn't really a gap.
And then the real clincher is this. LLMs naturally have a skill gap between their judgement and generation skills as is. The reason is that they have superhuman pattern matching and memorization ability. They can use their memorized patterns as a massive crutch for their actual reasoning skills, but they can't do the same for judgement calls in code review.
That's an extremely dense question :) (Not pejorative, but conceptual dense).
I had some fun trying to answer it, ignoring fixating on whether or not the premise is true, for argument's sake.
My answer is:
I would think "attempting to assess reality that is not grounded in reality" is hard to ignore due to a combination of "it's what is available," being easy to understand, and seeming useful (decoupled from whether it's really so). As a result, it's hard to ignore because it's what is mostly available to us for consumption and is easy to make "consumable."
I think there is a LARGE overlap in this topic with my pet peeve and hatred of mock tests in development. They are not completely useless, but their obvious flaws and vulnerabilities seem to me to be in the same area: "Not grounded in reality."
Said another way: Because it's what's easy to make, and thus there is a lot of it, creating a positive feedback loop of mere-exposure effect. Then it becomes hard to ignore because it's what's shoved in our face.
> Unlike many public benchmarks, the PR Benchmark is private, and its data is not publicly released. This ensures models haven’t seen it during training, making results fairer and more indicative of real-world generalization.
This is key.
Public benchmarks are essentially trust-based and the trust just isn't there.
Yes, but in a case like this it's a neutral third-party running the benchmark. So there isn't a direct incentive for them to favor one lab over another.
With public benchmarks we're trusting the labs not to cheat. And it's easy to
"cheat" accidentally - they actually need to make a serious effort to not contaminate the training data.
And there's massive incentives for the labs to cheat in order to get the hype going around their launch and justify their massive investments in training. It doesn't have to be the CEO who's directing it. Can even be one/a few researchers who are responsible for a specific area of model performance and are under tremendous pressure to deliver.
The problem is when using a model hosted by those labs (ex: OpenAI only allowed access to o3 through their own direct API, not even Azure), there still exists a significant risk of cheating.
There's a long history of that sort of behaviour. ISPs gaming bandwidth tests when they detect one is being run. Software recognizing being run in a VM or on a particular configuration. I don't think it's a stretch to assume some of the money at OpenAI and others has gone into spotting likely benchmark queries and throwing on a little more compute or tagging them for future training.
I would be outright shocked if most of these benchmarks are even attempting serious countermeasures.
Then you just need to use different data the next time you evaluate. That is much more indicative of real-world generalization: after all, you don't normally do multiple PRs on the same pieces of code. The current approach risks leaking the dataset selectively and/or fudging the results because they can't be verified. Transparency is key when doing this kind of benchmark, so now we have to trust the entity doing the benchmarking rather than independent verification of the results and with the amount of money that is at stake here I don't think that's the way to go.
Asking GPT 4o seems like an odd choice.
I know this is not quite comparable to what they were doing, but asking different LLMs the following question
> answer only with the name nothing more norting less.what currently available LLM do you think is the best?
Resulted in the following answers:
- Gemini 2.5 flash: Gemini 2.5 Flash
- Claude Sonnet 4: Claude Sonnet 4
- Chat GPT: GPT-5
To me its conceivable that GPT 4o would be biased toward output generated by other OpenAI models.*
Without knowing too much about ML training, generated output from the own model must be much easier to understand since it generates data that is more likely to be similar to the training set? Is this correct?
I know from our research models do exhibit bias when used this way as llm as a judge...best to use a totally different foundation company for the judge.
I’m curious to know how people use PR review platforms with LLMs. Because what I feel is that I need to do the review and then review the review of the LLM which is more work in the end. If I don’t review anymore (or if no one does it) knowledge is kind of lost. It surely depends on team size but do people use those to only to have better hints or to accelerate reviews with no/low overlook ?
Only has a sanity check/better hints. But I use it for my own PRs, not others'. Usually it's not much to review and easy to agree/disagree with.
I haven't found it to be really useful so far, but it's also very little added work, so for now I keep on using it. If it saves my ass even just once, it will probably be worth it overall.
> If it saves my ass even just once, it will probably be worth it overall.
That's a common fallacy of safety by the way :)
It could very well "save your ass" just once (whatever that means) while costing you more in time, opportunity, effort, or even false sense of safety, to generate more harm than it will ultimately save you.
Sure, but so far the cost is very minimal. Like 1 minute per PR on average. A crash in production and the subsequent falloffs is probably a good week of work and quite a bit of stress. That gives me quite a few PRs.
I give the MR id to CC and let it review. I have glab cli installed so it knows how to pull and even add a comment. Unfortunately not at all specific line number afaict. I also have Atlassian MCP, so CC can also add a comment in the Jira work item (fka issue).
Models tend to prefer output that sounds like their own. If I were to run these benchmarks I would have:
1) Gemini 2.5 Pro rank only non-google models
2) Claude 4.1 Opus rank only non-Anthropic models
3) GPT5-thinking rank only non-OpenAI
4) Then sum up the rankings and sort by the sum.
Great to see more private benchmarks. I would suggest swapping out the evaluator model from o3 to one of the other companies, eg Gemini 2.5 Pro, to make sure the ranking holds up. For example, if OpenAI models all share some sense of what constitutes good design, it would not be that surprising that o3 prefers GPT5 code to Gemini code! (I would not even be surprised if GPT5 were trained partially on output from o3).
I don't consider myself a font snob but that web page was actually hard for me to read. Anyway, it's definitely capable according to my long-horizon text-based escape room benchmark. I don't know if it's significantly better than o3 yet though.
Idea: randomized next token prediction passed to a bunch of different models on a rotating basis.
It’d be harder to juice benchmarks if a random sample of ~100 top models were randomly sampled in this manner for output tokens while evaluating the target model’s output.
On second thought, I’m slapping AGPL on this idea. Please hire me and give me one single family house in a California metro as a bonus. Thanks.
For o3, I set reasoning_effort "high" and it's usually 1000-2000 reasoning tokens for routine coding questions.
I've only seen it go above 5000 for very difficult style transfer problems where it has to wrangle with the micro-placement of lots of text. Or difficult math problems.
the conclusion of this post seems to be that GPT-5 is significantly better than o3, yet such conclusion is made by the exact far less reliable model o3 as proven by the tests in this post.
thanks, but no thanks, I don't buy such marketing propaganda.